The motivation of creating this notebook is a data science competition hosted by Microsoft and DrivenData, which is a part of the Capstone project of the Microsoft Professional Certificate in Data Science and also my very first competition related to machine learning. By the end of the competition, I was at the top 4% on the leaderboard, and I am happy to share with you my approach, especially to those who are just beginning their journey on data science.
Participants are to predict the probability that individuals across 7 different countries live below the poverty line at the $2.50/day threshold, given other socioeconomic indicators. The probability of being in poverty was calculated using the Poverty Probability Index (PPI), which estimates an individual's poverty status using 10 questions about a household’s characteristics and asset ownership. The remaining data comes from the Financial Inclusion Insights household surveys conducted by InterMedia.
The dataset contains the PPI along with 58 features of 12,600 individuals across 7 different countries.
I have hand-picked 3 regression-based models (Gboost, XGBoost and LightGBM) and used a 5-fold cross validation to evaluation their performance. A stacked model of the 3 is then tested for prediction results. The final r2 score was 0.4213, resulting a top 4% on the leaderboard.
The model stacking approach here is inspired by Serigne, make sure to check it out at: Serigne's Stacked Regressions Notebook.
Please check out my kaggle kernel for run result, and if you liked it - give it an up-vote! https://www.kaggle.com/johnnyyiu/poverty-prediction-from-visualization-to-stacking